sglang 0.4.4.post4__py3-none-any.whl → 0.4.5.post1__py3-none-any.whl

sglang/srt/models/mllama4.py

@@ -0,0 +1,227 @@
+from collections.abc import Iterable
+from typing import List, Optional, Set, Tuple
+
+import torch
+from torch import nn
+from transformers import Llama4Config, Llama4VisionModel
+from transformers.models.llama4.modeling_llama4 import Llama4MultiModalProjector
+
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
+from sglang.srt.layers.quantization import QuantizationConfig
+from sglang.srt.managers.mm_utils import (
+    MultiModalityDataPaddingPatternImageTokens,
+    general_mm_embed_routine,
+)
+from sglang.srt.managers.schedule_batch import MultimodalDataItem, MultimodalInputs
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_loader.weight_utils import default_weight_loader
+from sglang.srt.utils import add_prefix
+
+
+class Llama4ForConditionalGeneration(nn.Module):
+    packed_modules_mapping = {
+        "qkv_proj": ["q_proj", "k_proj", "v_proj"],
+        "gate_up_proj": ["gate_proj", "up_proj"],
+    }
+
+    def __init__(
+        self,
+        config: Llama4Config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.config = config
+        self.quant_config = quant_config
+
+        self.vision_model = Llama4VisionModel(config.vision_config)
+        self.multi_modal_projector = Llama4MultiModalProjector(config)
+
+        # Initialize the language model
+        from sglang.srt.models.llama4 import Llama4ForCausalLM
+
+        self.language_model = Llama4ForCausalLM(
+            config.text_config,
+            quant_config=quant_config,
+            prefix=add_prefix("language_model", prefix),
+        )
+
+        self.logits_processor = LogitsProcessor(config.text_config)
+
+    def pad_input_ids(self, input_ids: List[int], mm_inputs: MultimodalInputs):
+        # Get all special token IDs
+        im_token_id: int = mm_inputs.im_token_id
+
+        pattern = MultiModalityDataPaddingPatternImageTokens(torch.tensor(im_token_id))
+        return pattern.pad_input_tokens(input_ids, mm_inputs)
+
+    def get_image_feature(
+        self,
+        items: List[MultimodalDataItem],
+    ) -> torch.Tensor:
+        pixel_values = (
+            torch.concat([item.pixel_values for item in items])
+            .to(next(self.vision_model.parameters()).device)
+            .type(next(self.vision_model.parameters()).dtype)
+        )
+
+        image_outputs = self.vision_model(pixel_values, output_hidden_states=False)
+        image_features = image_outputs.last_hidden_state
+        vision_flat = image_features.view(-1, image_features.size(-1))
+        projected_vision_flat = self.multi_modal_projector(vision_flat)
+        return projected_vision_flat
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        **kwargs: object,
+    ) -> torch.Tensor:
+
+        hs = general_mm_embed_routine(
+            input_ids=input_ids,
+            forward_batch=forward_batch,
+            language_model=self.language_model,
+            image_data_embedding_func=self.get_image_feature,
+            positions=positions,
+        )
+
+        return hs
+
+    def permute_qk_weight_for_rotary(
+        self,
+        name: str,
+        loaded_weight: torch.Tensor,
+    ) -> Tuple[str, torch.Tensor]:
+
+        def permute(w: torch.Tensor, n_heads: int):
+            attn_in = self.language_model.config.head_dim * n_heads
+            attn_out = self.language_model.config.hidden_size
+
+            return (
+                w.view(n_heads, attn_in // n_heads // 2, 2, attn_out)
+                .transpose(1, 2)
+                .reshape(attn_in, attn_out)
+            )
+
+        modules = name.split(".")
+
+        # rotary embeds should be sliced
+        if ("wk" in modules or "k_proj" in modules) and modules[-1] == "weight":
+            loaded_weight = permute(
+                loaded_weight, self.language_model.config.num_key_value_heads
+            )
+        elif ("wq" in modules or "q_proj" in modules) and modules[-1] == "weight":
+            loaded_weight = permute(
+                loaded_weight, self.language_model.config.num_attention_heads
+            )
+
+        return name, loaded_weight
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]) -> Set[str]:
+
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            (".self_attn.qkv_proj", ".self_attn.q_proj", "q"),
+            (".self_attn.qkv_proj", ".self_attn.k_proj", "k"),
+            (".self_attn.qkv_proj", ".self_attn.v_proj", "v"),
+            (".shared_expert.gate_up_proj", ".shared_expert.gate_proj", 0),
+            (".shared_expert.gate_up_proj", ".shared_expert.up_proj", 1),
+            (".feed_forward.gate_up_proj", ".feed_forward.gate_proj", 0),
+            (".feed_forward.gate_up_proj", ".feed_forward.up_proj", 1),
+        ]
+
+        params_dict = dict(self.named_parameters())
+
+        num_experts = self.config.text_config.num_local_experts
+
+        # Params for weights, fp8 weight scales, fp8 activation scales
+        # (param_name, weight_name, expert_id, shard_id)
+        expert_params_mapping = FusedMoE.make_expert_params_mapping(
+            ckpt_gate_proj_name="gate_proj",
+            ckpt_down_proj_name="down_proj",
+            ckpt_up_proj_name="up_proj",
+            num_experts=num_experts,
+        )
+
+        for name, loaded_weight in weights:
+            if not "vision" in name:
+                name, loaded_weight = self.permute_qk_weight_for_rotary(
+                    name, loaded_weight
+                )
+
+            for param_name, weight_name, shard_id in stacked_params_mapping:
+                if weight_name not in name:
+                    continue
+
+                if "vision" in name:
+                    continue
+                name = name.replace(weight_name, param_name)
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                if ".experts" in name:
+                    # NOTE: llama4 fp8 has different weight format for experts
+                    if (
+                        "experts.gate_up_proj" not in name
+                        and "experts.down_proj" not in name
+                    ):
+                        for mapping in expert_params_mapping:
+                            param_name, weight_name, expert_id, shard_id = mapping
+                            if weight_name not in name:
+                                continue
+                            name = name.replace(weight_name, param_name)
+                            param = params_dict[name]
+                            weight_loader = param.weight_loader
+                            weight_loader(
+                                param,
+                                loaded_weight,
+                                name,
+                                shard_id=shard_id,
+                                expert_id=expert_id,
+                            )
+                            break
+                    else:
+                        if ".gate_up_proj" in name:
+                            name_list = [
+                                name.replace(
+                                    ".experts.gate_up_proj", ".experts.w13_weight"
+                                )
+                            ] * 2
+                            loaded_weight_list = loaded_weight.chunk(2, dim=-1)
+                            shard_id_list = ["w1", "w3"]
+                        else:
+                            name_list = [
+                                name.replace(".experts.down_proj", ".experts.w2_weight")
+                            ]
+                            shard_id_list = ["w2"]
+                            loaded_weight_list = [loaded_weight]
+                        for name, loaded_weight, shard_id in zip(
+                            name_list, loaded_weight_list, shard_id_list
+                        ):
+                            param = params_dict[name]
+                            weight_loader = param.weight_loader
+                            for expert_id in range(num_experts):
+                                weight_loader(
+                                    param,
+                                    loaded_weight[expert_id].T,
+                                    name,
+                                    shard_id=shard_id,
+                                    expert_id=expert_id,
+                                )
+                else:
+                    # Skip loading extra bias for GPTQ models.
+                    if name.endswith(".bias") and name not in params_dict:
+                        continue
+                    param = params_dict[name]
+                    weight_loader = getattr(
+                        param, "weight_loader", default_weight_loader
+                    )
+                    weight_loader(param, loaded_weight)
+
+
+EntryClass = Llama4ForConditionalGeneration

sglang/srt/models/olmo.py

@@ -93,6 +93,7 @@ class OlmoAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/olmo2.py

@@ -118,6 +118,7 @@ class Olmo2Attention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/olmoe.py

@@ -170,6 +170,7 @@ class OlmoeAttention(nn.Module):
             self.scaling,
             layer_id=layer_id,
             num_kv_heads=self.num_kv_heads,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/phi3_small.py

@@ -202,6 +202,7 @@ class Phi3SmallSelfAttention(nn.Module):
             self.scale,
             num_kv_heads=self.num_kv_heads_per_partion,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/qwen.py

@@ -133,6 +133,7 @@ class QWenAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/qwen2.py

@@ -154,6 +154,7 @@ class Qwen2Attention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/qwen2_5_vl.py

@@ -30,12 +30,16 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 from einops import rearrange
-from transformers import Qwen2VLConfig
 from transformers.activations import ACT2FN
 from transformers.models.qwen2.modeling_qwen2 import Qwen2RMSNorm
 from transformers.models.qwen2_5_vl.configuration_qwen2_5_vl import (
+    Qwen2_5_VLConfig,
     Qwen2_5_VLVisionConfig,
 )
+from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
+    Qwen2_5_VisionPatchEmbed,
+    Qwen2_5_VisionRotaryEmbedding,
+)
 
 from sglang.srt.hf_transformers_utils import get_processor
 from sglang.srt.layers.attention.vision import VisionAttention
@@ -137,7 +141,7 @@ class Qwen2_5_VisionBlock(nn.Module):
             embed_dim=dim,
             num_heads=num_heads,
             projection_size=dim,
-            use_qkv_parallel=False,
+            use_qkv_parallel=True,
             use_context_forward=use_context_forward,
             softmax_in_single_precision=softmax_in_single_precision,
             flatten_batch=flatten_batch,
@@ -173,33 +177,6 @@ class Qwen2_5_VisionBlock(nn.Module):
         return x
 
 
-class Qwen2_5_VisionPatchEmbed(nn.Module):
-
-    def __init__(
-        self,
-        patch_size: int = 14,
-        temporal_patch_size: int = 2,
-        in_chans: int = 3,
-        embed_dim: int = 1152,
-    ) -> None:
-        super().__init__()
-        self.patch_size = patch_size
-        self.temporal_patch_size = temporal_patch_size
-        self.embed_dim = embed_dim
-
-        kernel_size = [temporal_patch_size, patch_size, patch_size]
-        self.proj = nn.Conv3d(
-            in_chans, embed_dim, kernel_size=kernel_size, stride=kernel_size, bias=False
-        )
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        target_dtype = self.proj.weight.dtype
-        L, C = x.shape
-        x = x.view(L, -1, self.temporal_patch_size, self.patch_size, self.patch_size)
-        x = self.proj(x.to(dtype=target_dtype)).view(L, self.embed_dim)
-        return x
-
-
 class Qwen2_5_VisionPatchMerger(nn.Module):
 
     def __init__(
@@ -244,21 +221,6 @@ class Qwen2_5_VisionPatchMerger(nn.Module):
         return out
 
 
-class Qwen2_5_VisionRotaryEmbedding(nn.Module):
-
-    def __init__(self, dim: int, theta: float = 10000.0) -> None:
-        super().__init__()
-        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
-        self.register_buffer("inv_freq", inv_freq, persistent=False)
-
-    def forward(self, seqlen: int) -> torch.Tensor:
-        seq = torch.arange(
-            seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype
-        )
-        freqs = torch.outer(seq, self.inv_freq)
-        return freqs
-
-
 class Qwen2_5_VisionTransformer(nn.Module):
 
     def __init__(
@@ -275,7 +237,7 @@ class Qwen2_5_VisionTransformer(nn.Module):
         spatial_merge_size: int = vision_config.spatial_merge_size
         self.spatial_merge_size = spatial_merge_size
         self.spatial_merge_unit: int = spatial_merge_size * spatial_merge_size
-        in_chans: int = vision_config.in_channels
+        in_channels: int = vision_config.in_channels
         hidden_size: int = vision_config.hidden_size
         depth: int = vision_config.depth
         num_heads: int = vision_config.num_heads
@@ -286,7 +248,7 @@ class Qwen2_5_VisionTransformer(nn.Module):
         self.patch_embed = Qwen2_5_VisionPatchEmbed(
             patch_size=patch_size,
             temporal_patch_size=temporal_patch_size,
-            in_chans=in_chans,
+            in_channels=in_channels,
             embed_dim=hidden_size,
         )
 
@@ -363,7 +325,7 @@ class Qwen2_5_VisionTransformer(nn.Module):
 
     @property
     def dtype(self) -> torch.dtype:
-        return self.blocks[0].mlp.gate_proj.weight.dtype
+        return self.patch_embed.proj.weight.dtype
 
     @property
     def device(self) -> torch.device:
@@ -467,9 +429,28 @@ cached_get_processor = lru_cache(get_processor)
 
 
 class Qwen2_5_VLForConditionalGeneration(nn.Module):
+    # BitandBytes specific attributes
+    default_bitsandbytes_target_modules = [
+        ".gate_proj.",
+        ".down_proj.",
+        ".up_proj.",
+        ".q_proj.",
+        ".k_proj.",
+        ".v_proj.",
+        ".o_proj.",
+    ]
+    bitsandbytes_stacked_params_mapping = {
+        # shard_name, weight_name, index
+        "q_proj": ("qkv_proj", 0),
+        "k_proj": ("qkv_proj", 1),
+        "v_proj": ("qkv_proj", 2),
+        "gate_proj": ("gate_up_proj", 0),
+        "up_proj": ("gate_up_proj", 1),
+    }
+
     def __init__(
         self,
-        config: Qwen2VLConfig,
+        config: Qwen2_5_VLConfig,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
     ) -> None:
@@ -479,9 +460,9 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
         self.visual = Qwen2_5_VisionTransformer(
             config.vision_config,
             norm_eps=getattr(config, "rms_norm_eps", 1e-6),
-            # NOTE: Qwen2-VL vision encoder does not support any
-            # quantization method now.
-            quant_config=None,
+            # NOTE: Qwen2_5-VL vision encoder currently supports BitsAndBytes 4-bit quantization.
+            # Other quantization methods (e.g., GPTQ, AWQ) are untested and may not be supported.
+            quant_config=quant_config,
             prefix=add_prefix("visual", prefix),
         )
 
@@ -500,6 +481,7 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
                 quant_config=quant_config,
                 prefix=add_prefix("lm_head", prefix),
             )
+        self.is_mrope_enabled = "mrope_section" in self.config.rope_scaling
 
         self.logits_processor = LogitsProcessor(config)
         self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
@@ -553,14 +535,14 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
                 otherwise it will be `(seq_len,).
                 (Use input_metadata.mrope_positions to replace it)
         """
-        if getattr(self.config, "rope_scaling", {}).get("type", None) == "mrope":
+        if self.is_mrope_enabled:
             positions = forward_batch.mrope_positions
 
         if not (
             forward_batch.forward_mode.is_decode()
             or not forward_batch.contains_image_inputs()
         ):
-            if getattr(self.config, "rope_scaling", {}).get("type", None) == "mrope":
+            if self.is_mrope_enabled:
                 assert positions.ndim == 2 and positions.size(0) == 3, (
                     "multimodal section rotary embedding requires "
                     f"(3, seq_len) positions, but got {positions.size()}"
@@ -610,23 +592,6 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
                 weight_loader(param, loaded_weight, shard_id)
                 break
             else:
-                if "visual" in name and "qkv.weight" in name:
-                    visual_num_heads = self.config.vision_config.num_heads
-                    visual_embed_dim = self.config.vision_config.hidden_size
-                    head_size = visual_embed_dim // visual_num_heads
-                    loaded_weight = loaded_weight.view(
-                        3, visual_num_heads, head_size, visual_embed_dim
-                    )
-                    loaded_weight = loaded_weight.transpose(0, 1)
-                    loaded_weight = loaded_weight.reshape(-1, visual_embed_dim)
-                elif "visual" in name and "qkv.bias" in name:
-                    visual_num_heads = self.config.vision_config.num_heads
-                    visual_embed_dim = self.config.vision_config.hidden_size
-                    head_size = visual_embed_dim // visual_num_heads
-                    loaded_weight = loaded_weight.view(3, visual_num_heads, head_size)
-                    loaded_weight = loaded_weight.transpose(0, 1)
-                    loaded_weight = loaded_weight.reshape(-1)
-
                 if "visual" in name:
                     # adapt to VisionAttention
                     name = name.replace(r"attn.qkv.", r"attn.qkv_proj.")

sglang/srt/models/qwen2_moe.py

@@ -231,6 +231,7 @@ class Qwen2MoeAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/qwen2_vl.py

@@ -152,7 +152,7 @@ class Qwen2VisionBlock(nn.Module):
             embed_dim=dim,
             num_heads=num_heads,
             projection_size=dim,
-            use_qkv_parallel=False,
+            use_qkv_parallel=True,
             use_context_forward=use_context_forward,
             softmax_in_single_precision=softmax_in_single_precision,
             flatten_batch=True,
@@ -351,7 +351,7 @@ class Qwen2VisionTransformer(nn.Module):
 
     @property
     def dtype(self) -> torch.dtype:
-        return next(self.parameters()).dtype
+        return self.patch_embed.proj.weight.dtype
 
     @property
     def device(self) -> torch.device:
@@ -423,6 +423,25 @@ cached_get_processor = lru_cache(get_processor)
 
 
 class Qwen2VLForConditionalGeneration(nn.Module):
+    # BitandBytes specific attributes
+    default_bitsandbytes_target_modules = [
+        ".gate_proj.",
+        ".down_proj.",
+        ".up_proj.",
+        ".q_proj.",
+        ".k_proj.",
+        ".v_proj.",
+        ".o_proj.",
+    ]
+    bitsandbytes_stacked_params_mapping = {
+        # shard_name, weight_name, index
+        "q_proj": ("qkv_proj", 0),
+        "k_proj": ("qkv_proj", 1),
+        "v_proj": ("qkv_proj", 2),
+        "gate_proj": ("gate_up_proj", 0),
+        "up_proj": ("gate_up_proj", 1),
+    }
+
     def calculate_num_image_tokens(self, image_grid_thw: Tuple[int, int, int]):
         processor = cached_get_processor(self.config._name_or_path)
         grid_t, grid_h, grid_w = image_grid_thw
@@ -447,9 +466,9 @@ class Qwen2VLForConditionalGeneration(nn.Module):
         self.visual = Qwen2VisionTransformer(
             config.vision_config,
             norm_eps=getattr(config, "rms_norm_eps", 1e-6),
-            # NOTE: Qwen2-VL vision encoder does not support any
-            # quantization method now.
-            quant_config=None,
+            # NOTE: Qwen2-VL vision encoder currently supports BitsAndBytes 4-bit quantization.
+            # Other quantization methods (e.g., GPTQ, AWQ) are untested and may not be supported.
+            quant_config=quant_config,
             prefix=add_prefix("visual", prefix),
         )
 
@@ -467,6 +486,7 @@ class Qwen2VLForConditionalGeneration(nn.Module):
                 prefix=add_prefix("lm_head", prefix),
             )
 
+        self.is_mrope_enabled = "mrope_section" in self.config.rope_scaling
         self.logits_processor = LogitsProcessor(config)
         self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
 
@@ -521,14 +541,14 @@ class Qwen2VLForConditionalGeneration(nn.Module):
                 otherwise it will be `(seq_len,).
                 (Use input_metadata.mrope_positions to replace it)
         """
-        if getattr(self.config, "rope_scaling", {}).get("type", None) == "mrope":
+        if self.is_mrope_enabled:
             positions = forward_batch.mrope_positions
 
         if not (
             forward_batch.forward_mode.is_decode()
             or not forward_batch.contains_image_inputs()
         ):
-            if getattr(self.config, "rope_scaling", {}).get("type", None) == "mrope":
+            if self.is_mrope_enabled:
                 assert positions.ndim == 2 and positions.size(0) == 3, (
                     "multimodal section rotary embedding requires "
                     f"(3, seq_len) positions, but got {positions.size()}"
@@ -577,24 +597,6 @@ class Qwen2VLForConditionalGeneration(nn.Module):
                 weight_loader(param, loaded_weight, shard_id)
                 break
             else:
-
-                if "visual" in name and "qkv.weight" in name:
-                    visual_num_heads = self.config.vision_config.num_heads
-                    visual_embed_dim = self.config.vision_config.embed_dim
-                    head_size = visual_embed_dim // visual_num_heads
-                    loaded_weight = loaded_weight.view(
-                        3, visual_num_heads, head_size, visual_embed_dim
-                    )
-                    loaded_weight = loaded_weight.transpose(0, 1)
-                    loaded_weight = loaded_weight.reshape(-1, visual_embed_dim)
-                elif "visual" in name and "qkv.bias" in name:
-                    visual_num_heads = self.config.vision_config.num_heads
-                    visual_embed_dim = self.config.vision_config.embed_dim
-                    head_size = visual_embed_dim // visual_num_heads
-                    loaded_weight = loaded_weight.view(3, visual_num_heads, head_size)
-                    loaded_weight = loaded_weight.transpose(0, 1)
-                    loaded_weight = loaded_weight.reshape(-1)
-
                 if "visual" in name:
                     # adapt to VisionAttention
                     name = name.replace(r"attn.qkv.", r"attn.qkv_proj.")

sglang/srt/models/stablelm.py

@@ -149,6 +149,7 @@ class StablelmAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_key_value_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/xverse.py

@@ -153,6 +153,7 @@ class XverseAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/models/xverse_moe.py

@@ -252,6 +252,7 @@ class XverseAttention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_id,
+            quant_config=quant_config,
             prefix=add_prefix("attn", prefix),
         )
 

sglang/srt/openai_api/adapter.py

@@ -983,6 +983,8 @@ def v1_chat_generate_request(
                     ):
                         encoded = encoded[1:]
                     prompt_ids += encoded
+                if tokenizer_manager.model_config.is_multimodal:
+                    prompt = tokenizer_manager.tokenizer.decode(prompt_ids)
                 stop = request.stop
                 image_data = None
                 audio_data = None
@@ -993,7 +995,8 @@ def v1_chat_generate_request(
                 image_data = conv.image_data
                 audio_data = conv.audio_data
                 modalities = conv.modalities
-                stop = conv.stop_str or []
+                stop = conv.stop_str or [] if not request.ignore_eos else []
+
                 if request.stop:
                     if isinstance(request.stop, str):
                         stop.append(request.stop)

sglang/srt/patch_torch.py

@@ -14,6 +14,7 @@
 from typing import Callable, Union
 
 import torch
+from packaging import version
 from torch.multiprocessing import reductions
 
 
@@ -69,3 +70,13 @@ def _device_from_maybe_uuid(device_maybe_uuid: Union[int, str]) -> int:
 
 def _modify_tuple(t, index: int, modifier: Callable):
     return *t[:index], modifier(t[index]), *t[index + 1 :]
+
+
+def monkey_patch_torch_compile():
+    if version.parse(torch.__version__) < version.parse("2.8.0"):
+        # These things are cacheable by torch.compile. torch.compile just doesn't know it.
+        # This was fixed in PyTorch 2.8, but until then, we monkey patch.
+        import torch._higher_order_ops.auto_functionalize as af
+
+        af.auto_functionalized_v2._cacheable = True
+        af.auto_functionalized._cacheable = True